Improving the Automatic Classification of Brain MRI Acquisition Contrast with Machine Learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Automated quantification of MRI data requires accurate identification of specific acquisition types (e.g., T1-weighted, T2-weighted).
• Previous methods using imaging or DICOM metadata showed success in controlled settings.
• Heterogeneous clinical datasets from diverse scanners and institutions present a challenge for reliable classification.

Purpose of the Study:

• To compare the performance of imaging-based and metadata-based methods for classifying MRI acquisition contrast types.
• To develop and evaluate robust models for accurate MRI contrast type classification across varied clinical datasets.
• To assess the generalizability of classification methods for high-throughput medical image analysis.

Main Methods:

• Developed and trained Random Forest (RF) and Convolutional Neural Network (CNN) models using DICOM metadata and pixel data, respectively.
• Created a combined RF model integrating CNN outputs with metadata features.
• Evaluated models on four diverse cohorts: MS research, MS clinical, glioma research, and ADNI PTSD, encompassing various scanners and pathologies.

Main Results:

• Pixel-based CNN and combined models achieved 97-98% accuracy on the clinical MS cohort.
• The metadata-only model reached 99.7% accuracy, while the CNN model achieved 98.4% on the glioma test/validation cohort.
• Demonstrated accurate and generalizable classification of MRI acquisition contrast types across heterogeneous datasets.

Conclusions:

• Accurate and generalizable classification of MRI acquisition contrast types is feasible using machine learning approaches.
• These methods are vital for automating data selection in high-throughput and big-data medical image analysis.
• The developed models show promise for improving efficiency and reliability in clinical research and practice.